Re: Intel QAT on A2SDi-8C-HLN4F causes massive data corruption with dm-crypt + xfs

From: Eric Biggers <ebiggers@kernel.org>
To: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>,
	Greg KH <gregkh@linuxfoundation.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Kyle Sanderson <kyle.leet@gmail.com>,
	Dave Chinner <david@fromorbit.com>,
	qat-linux@intel.com, Linux-Kernal <linux-kernel@vger.kernel.org>,
	linux-xfs <linux-xfs@vger.kernel.org>,
	Linux Crypto Mailing List <linux-crypto@vger.kernel.org>,
	device-mapper development <dm-devel@redhat.com>
Subject: Re: Intel QAT on A2SDi-8C-HLN4F causes massive data corruption with dm-crypt + xfs
Date: Thu, 3 Mar 2022 21:44:53 +0000	[thread overview]
Message-ID: <YiE21aPGG5DHwJvb@gmail.com> (raw)
In-Reply-To: <YiEyGoHacN80FcOL@silpixa00400314>

On Thu, Mar 03, 2022 at 09:24:42PM +0000, Giovanni Cabiddu wrote:
> On Thu, Mar 03, 2022 at 07:21:33PM +0000, Eric Biggers wrote:
> > If these algorithms have critical bugs, which it appears they do, then IMO it
> > would be better to disable them (either stop registering them, or disable the
> > whole driver) than to leave them available with low cra_priority.  Low
> > cra_priority doesn't guarantee that they aren't used.
> Thanks for your feedback Eric.
> 
> Here is a patch that disables the registration of the algorithms in the
> QAT driver by setting, a config time, the number of HW queues (aka
> instances) to zero.
> 
> ---8<---
> From: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
> Subject: [PATCH] crypto: qat - disable registration of algorithms
> Organization: Intel Research and Development Ireland Ltd - Co. Reg. #308263 - Collinstown Industrial Park, Leixlip, County Kildare - Ireland
> 
> The implementations of aead and skcipher in the QAT driver do not
> support properly requests with the CRYPTO_TFM_REQ_MAY_BACKLOG flag set.
> If the HW queue is full, the driver returns -EBUSY but does not enqueue
> the request.
> This can result in applications like dm-crypt waiting indefinitely for a
> completion of a request that was never submitted to the hardware.
> 
> To avoid this problem, disable the registration of all skcipher and aead
> implementations in the QAT driver by setting the number of crypto
> instances to 0 at configuration time.
> 
> This patch deviates from the original upstream solution, that prevents
> dm-crypt to use drivers registered with the flag
> CRYPTO_ALG_ALLOCATES_MEMORY, since a backport of that set to stable
> kernels may have a too wide effect.
> 
> commit 7bcb2c99f8ed032cfb3f5596b4dccac6b1f501df upstream
> commit 2eb27c11937ee9984c04b75d213a737291c5f58c upstream
> commit fbb6cda44190d72aa5199d728797aabc6d2ed816 upstream
> commit b8aa7dc5c7535f9abfca4bceb0ade9ee10cf5f54 upstream
> commit cd74693870fb748d812867ba49af733d689a3604 upstream
> 
> Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
> ---
>  drivers/crypto/qat/qat_common/qat_crypto.c | 4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)

Sounds good; is there any reason not to apply this upstream too, though?
You could revert it later as part of the patch series that fixes the driver.

- Eric